Delineating contrast enhancing (CE) tissue is an integral part of the RANO criteria for therapy response assessment in high-grade gliomas. Manual delineations of brain tumours provide an approximate separation between the CE region of interest (ROI) and perilesional edema (ED) ROI. The focus of this study is to evaluate the impact of semi-automatic delineation of hotspots of CE (HCE) in brain tumour follow-up of glioblastoma multiforme (GBM) patients after surgery. Based on multi-parametric magnetic resonance data we predict the post-operative evolution of the brain tumour by labelling each patient at each time point as responsive or progressive. We compare the results obtained with our semi-automatic method to the results obtained with the original manual delineations.

Acquisition

Twenty-nine GBM patients who underwent surgery were scanned monthly using a 3 Tesla MRI unit (Philips Achieva, Best, The Netherlands). One scanning session consisted of conventional MRI (T1-weighted MRI before and after contrast administration, T2-weighted MRI and FLAIR (fluid attenuated inversion recovery) MRI) and advanced MRI (diffusion kurtosis imaging (DKI), dynamic-susceptibility weighted contrast (DSC) - MRI). ROIs were manually drawn on the T1-post contrast (T1pc) images by an expert radiologist around the solid CE region and the entire lesion (TO), i.e. CE and ED. Another ROI was drawn around the contralateral normal appearing white matter (NAWM) to standardize the hemodynamic measurements of DSC-MRI ¹. A label (responsive or progressive) has been put on each patient at a certain time point (TP0) according to the RANO criteria.

Proposed method

We create another set of ROIs in the following way: for each session of each patient we take the T1pc intensities of all manually delineated voxels (CE & ED) and set an intensity threshold at the 90th percentile (P90). All the voxels with intensities higher than or equal to P90 belong to the new HCE, while the rest are labelled as extended ED (EED).

Feature extraction

After processing DSC-MRI and DKI-MRI we have 8 perfusion maps and 7 diffusion maps, respectively. To these we add the four conventional maps, so in total there are 19 parameter maps. For each patient session we perform an affine co-registration of the parameter maps to the T1pc map. For each parameter map we normalize the intensities in each tumour ROI (CE, ED, HCE, EED) to the average value of the voxel intensities in the NAWM ROI. Within each tumour ROI (CE, ED, HCE, EED) we compute the average, coefficient of variation, kurtosis, skewness, 10th percentile and 90th percentile of the voxel intensities for each parameter map. In total, from 29 patients, we have 43 data points, each with 228 features. The 43 data points correspond to a timeframe between TP0 and 6 months after TP0.

We perform feature selection on each set of ROIs (CE+ED vs. HCE+EED) by learning Random Forest (RF) feature relevance on the training sets. Because we want to study the impact of our new delineations compared to the manual delineations, we select a common set of five features that have high relevance for the two sets of ROIs (table 1).

By using our approach as an imputation method, we add 12 points to the dataset, using only CE or ED as starting delineation, not the reunion CE&ED, as described in Methods. We keep the same 5 features. We test 4 classifiers (Support Vector Machines (SVM), Linear Discriminant Analysis (LDA), RUSBoost and Bagging) using Leave One Patient Out Cross Validation (LOPO-CV). We measure performance by computing the balanced error rate (BER) for each time point and the weighted BER (wBER) for all time points. ²

Automatic feature selection revealed that (H)CE provides the most important features. The classification results using our semi-automatic delineations are better in almost all cases (table 2) compared to those using the manual delineations. Most of the 12 added points did not have a manual CE ROI because the expert did not see any enhancing tissue, most likely due to the fact that all 12 points are labelled as responsive. After imputing features using our approach, we notice a significant decrease in wBER for all classifiers. This is mostly due to a better learning stage of the classifiers, because now the responsive class has 12 more data points (table 3).Our new semi-automatic delineation method provides a better classification between responsive and progressive patients and can efficiently be used to impute missing features.